Human force reproduction error depends upon force levelBram Onneweer∗

Delft University of TechnologyWinfred Mugge†

Delft University of TechnologyAlfred C. Schouten‡

Delft University of TechnologyTwente University

ABSTRACT

For optimal haptic tele-manipulation system design, it is importantto understand the accuracy and limitations of human force percep-tion. Previous research demonstrated that humans generate higherforces when asked to reproduce an externally applied force; thesestudies proposed that the nervous system attenuates feedback fromself-generated forces. The goal of this study was to determine howaccurately subjects reproduce self-generated forces with the samehand over a broad range of force levels. Subjects (n=10, all righthanded) had to generate an onscreen target force with visual sup-port and subsequently reproduce the same force without visual sup-port with their right hand against a static handle equipped with aforce sensor. Six force levels (10 to 160N) were each presentedrandomly for eight repetitions. Subjects generated too high forcesfor lower force levels (≤ 40N) and too low forces for higher forcelevels (≥ 130N). Our results support force-dependent sensory inte-gration and demonstrate that attenuated feedback of self-generatedforces is not the sole factor in force reproduction errors.

Keywords: Attenuation, Force reproduction error, self-generatedforces, haptic tele-manipulation systems, sensory integration

1 INTRODUCTION

In situations where human judgement is wanted, but human pres-ence is undesirable such as in hazardous environments (e.g. nuclearpower plants), haptic tele-manipulation systems provide the solu-tion. The human operator controls a slave (e.g. remote robot arm),via the master (e.g. joystick) while haptic information of the forcesat the slave is fed back to the human operator. Yet, the delicate na-ture of most tele-manipulation tasks complicates proper executionusing tele-manipulation systems, leading to accidents. The accu-racy and limitations of human force perception is a critical factor inoptimizing man-machine interaction.

Humans perceive forces using Golgi Tendon Organs (GTO) thatdetect forces in the muscles and tactile sensors that detect defor-mations of the skin due to forces. When multiple sensors pro-vide redundant information, the central nervous system integratesthe information, using the accuracy of the sensorsy information asa weighting factor [3], [4], [6], [8]. To interact with the environ-ment, humans generate forces using their muscles. Next to the ex-ternally applied forces, also self-generated forces are detected bythe force sensors and will affect the external force estimate. Theefference copy (motor commands used to generate forces with theown body), provides additional information that might correct forthis extra sensory information [1] or can be used in the integrationprocess to improve the force estimate.

In literature, it has been demonstrated that humans generatehigher forces at the fingertip when asked to reproduce an externallyapplied target force up to 10N [5], [7]. It has been proposed that the

∗e-mail: b.onneweer@tudelft.nl†e-mail: w.mugge@tudelft.nl‡e-mail: a.c.schouten@tudelft.nl

nervous system attenuates feedback from self-generated forces, i.e.humans perceive self-generated forces to be lower than externallyapplied forces [5]. Bays and Wolpert [1], proposed a model that,using the efference copy, predicts the reafference feedback (sensoryfeedback due to self-generated forces), which is attenuated by sub-tracting it from the actual sensory feedback.

Walsh et al. [7] suggest that the force reproduction error (FRE,force difference between target and, in a subsequent trial, repro-duced force) depends on a constant component (i.e. offset) and aforce-level-dependent component (i.e. gradient). The FRE dimin-ished when subjects self-generated the target force using their otherhand, which is in accordance with Bays and Wolpert’s model, butalso diminished when subjects matched higher externally appliedforce levels (up to 75 % of their maximum voluntary contraction),which is not captured by the model as concluded by Walsh [7].Walsh et al. [7] proposed that the gradient is generated due to anenhancement of reafference when receiving the target force, i.e.humans are more sensitive to self-generated reaction forces thanto the externally applied force. Jones and Hunter [2], demonstratedthat subjects generated relatively higher forces for low force levelswith diminished FREs for high force levels when matching isomet-ric contraction levels of the other arm on the basis of equal sensa-tion. These previous studies present the FRE when matching ex-ternally generated targets or when targets are self-generated by theother arm. In tele-manipulations systems, however, humans oftenhave to reproduce forces in the same way as performed before (e.g.in training sessions).

The goal of our study was to determine how accurately subjectsreproduce self-generated forces over a broad range of force lev-els, from 10N to 160N, with the same hand. If only attenuation ofself-generated forces causes the FRE, then self-generating both thetarget and the reproduction force should result in correct force esti-mates and no FRE. Therefore, attenuation of self-generated forcesas proposed by Bays and Wolpert [1] does not affect the FRE in ourexperiment. We hypothesize that the accuracies of the force per-ception sensors change with force level and that subsequent sensoryreweighting will affect the FRE. We expect that subjects will gen-erate too high forces for force levels near 10N and too low forcesfor force levels near 160N. The results of this study will indicatewhether attenuation of self-generated forces is the sole factor inforce reproduction tasks. In addition, the results may have an im-pact on the design of haptic tele-manipulation systems. Knowledgeof the human force perception can be utilized in scaling environ-mental forces to and from the human operator, e.g. preventing dam-age to the hazardous environment by presenting the environmentalforce in a range where the operator’s FRE is smallest.

2 MATERIALS AND METHODS

Subjects Ten healthy right-handed males, aged 27.1 (SD1.37), participated in the experiment. All subjects provided writteninformed consent prior to participation and the study was approvedby the local ethics committee.

617

IEEE World Haptics Conference 201314-18 April, Daejeon, Korea978-1-4799-0088-6/13/$31.00 ©2013 IEEE

Experimental setup Figure 1 shows the experimental setupwith a subject in front of a monitor exerting forces to a handle withits right hand. Subjects performed two types of trials:

• a reference trial, where the subject generated an onscreen tar-get force with visual support

• a reproduction trial, where the force of the reference trial wasreproduced without visual support

The visual support consisted of target indicators and an indica-tor bar, see Fig. 1. The subjects were instructed to operate a footswitch that triggered a force measurement (0.3s with a sample rateof 500 Hz) when the bar was aligned with the target indicators (ref-erence trial) or when they thought to have matched the referenceforce (reproduction trial). After operating the foot switch, subjectswere instructed to maintain the force for at least one second (timedonscreen) to prevent the force to change during the measurement.

Experimental protocol The maximum voluntary contractions(MVC) before and after the experiment were used to check for fa-tigue. To prevent impact during the MVC trials, subjects were in-structed to build up to their maximum force in three seconds. Visualfeedback of the applied force as well as the maximum force fromthe previous trial was shown. The maximum force per trial was cal-culated using a moving average filter (time window: 100ms). Threetrials were performed and the maximum force of the trials were av-eraged to obtain the MVC.

The experiment started with a training session to familiarize thesubject with the setup and the protocol and consisted of five refer-ence trials followed by eight alternating reference and reproductiontrials at a force level of 20N. The next trial started when the forceon the handle was back to zero. Subjects were instructed not to letgo of the handle between a reference and reproduction trial.

In the experiment, the subjects had to perform alternating refer-ence and reproduction trials of six force levels (10, 40, 70, 100, 130,160N), which were randomly presented eight times each. Resultingin a total of 96 trials (6 force level x 2 trial type x 8 repetitions.

Force

Force sensor

Foot switch

Figure 1: Experimental setup. The subjects were seated in front ofa monitor and held a handle, equipped with a force sensor (ATI mini45, calibration SI-145-5, sensitivity 1/8N), in their right hand with theirright arm at an elbow angle of approximately 90 degrees. Visual sup-port consisted of an indicator bar and target indicators. The indicatorbar showed the force applied by the subject scaled to the target force,resulting in the same visual support for all force levels with the targetindicators at 100%.

Data analysis The force measurements during the referenceand reproduction trials were averaged over the 0.3 s measurementand the repetitions. To obtain the FRE, the target force is subtractedfrom the averaged reproduced force and the relative FRE is obtainedby dividing the FRE by the corresponding target force. Repeated-measures ANOVA was used to test for effects of force level on theFRE and on the relative FRE, with α = 0.05.

To check if the FRE and relative FRE differed significantly fromzero, separate one way t-tests were performed. The α was correctedusing Bonferroni correction. To measure the effect of fatigue, apaired t-test was performed between the MVC before and after theexperiment. Statistical analysis was performed in IBM SPSS Statis-tics 20.

3 RESULTS

The MVC results from Table 1 indicate that there was no effect offatigue on the results (t(9)= 0.014, p = 0.989).

The small variance within subjects demonstrate that the taskswas repeatable, see Fig. 2A.

A

0

20

40

60

80

100

120

140

160

Rep

rodu

ced

Forc

e [N

]

0 20 40 60 80 100 120 140 160Target Force [N]

B

0 20 40 60 80 100 120 140 1600

20

40

60

80

100

120

140

160

Target Force [N]

Rep

rodu

ced

Forc

e [N

]

Figure 2: A, reproduced forces of a typical subject. Dashed blackline: target forces; Errorbars: Standard error across trials. B, Aver-aged reproduction forces of all subjects. Dashed black line: targetforces; Markers: different subjects. Nine out of ten subjects, gener-ated higher forces for force levels below 40N and generated lowerforces for force levels of 130N and up.

618

Table 1: Maximum voluntary contraction per subject and group aver-age. MVC B: before experiment; MVC A: after experiment; Diff [N]:MVC A - MVC B; Diff [%] = Diff [N] / MVC B.

Subject MVC B [N] MVC A [N] Diff. [N] Diff [%]1 321.4 323.0 1.6 0.52 338.6 300.9 -37.7 -12.53 304.0 260.3 -43.7 -16.84 372.8 402.8 30.0 7.55 368.8 369.3 0.5 0.16 279.9 280.0 0.1 0.17 422.4 409.1 -13.3 -3.28 298.0 370.3 72.3 19.59 382.3 302.3 -80.0 -26.510 340.9 409.0 68.1 16.7

mean 342.9 342.7 -0.20 0.06

Table 2: Force reproduction error. Per force level the mean FRE,standard error of FRE across subjects, mean relative FRE and stan-dard error of the relative FRE across subjects; p-value: significancelevel of one way t-test. The α of the t-test is Bonferroni corrected to0.008.

Force [N] Mean [N] SE [N] Mean [%] SE [%] p-value10 6.52 1.59 65.2 15.9 0.00240 9.74 1.98 23.6 4.9 0.00170 3.44 2.57 4.5 3.7 0.198

100 -5.32 1.81 -5.9 1.8 0.015130 -13.59 2.50 -11.2 1.9 <0.001160 -20.88 2.49 -13.8 1.6 <0.001

The variance of the force estimates between subjects is largestfor 130N and 160N. The results show that there is an effect of forcelevel on the FRE (F5,45=76.739, p<0.001), see Fig. 3A. Thereis also an effect of force level on the relative FRE (F5,45=24.655,p<0.001), see Fig. 3B. Subjects generated too high forces for forcelevels up to 40N (10N: p=0.002, 40N: p=0.001) and too low forcesfor force level above 130 N (130N and 160N: p<0.001), with thecrossover point between 40N and 130N see Fig.3A and table 2. At10N, the relative FRE and the range of force estimates between sub-jects is largest, see Fig.3B, for higher force levels, 40N until 160N,the relative FRE and the variance between subjects decreases.

4 DISCUSSION

In this study we analysed the effect of force level on the force re-production error when reproducing self-generated forces using thesame hand. As hypothesized, an effect of force level on FRE wasfound. Subjects generated too high forces for force levels up to 40Nand too low forces for force levels of 130N and up. The cross-overpoint was comparable for all subjects, see Fig. 2B, suggesting thatthe MVC does not affect the location of the cross-over point. Thegroup average MVC results demonstrate that there was no fatigueduring this experiment.

Force reproduction error Our results for low force levels (upto 40N) are in accordance with previous studies where subjects gen-erated too high forces at the fingertip when matching externally ap-plied target forces up to 10N [5], [7]. However, our findings whenusing the same hand are not in accordance with findings where theFRE diminished when subjects had to self-generate the target force,up to 10N, using their other hand [7]. Apparently human force per-ception depends on the way the forces are experienced, i.e. which

sensors obtain useful information to be used in the reproductionphase.

Subjects generated too low forces for force levels of 130N andup, which is not in accordance with previous findings where theFRE diminished at higher externally applied force levels (up to75% MVC) [7] and when matching higher contraction levels (upto 85% MVC) of the biceps muscle of the other arm [2]. It wasfound that subjects generated too low forces at higher force levels,when only tactile sensors were used to perceive the externally ap-plied force [7], indicating that the CNS uses information from allavailable sensors to estimate the applied force.

Possible mechanisms Up to now, the discussion focused onthe FRE, but more important is the mechanism behind the FRE.The model of Bays and Wolpert [1] proposed that the CNS attenu-ates self-generated feedback to estimate the sensory feedback dueto external influences. In this study, both the target and reproduc-tion force are self-generated using the same hand. Attenuation ofself-generated feedback as proposed by Bays and Wolpert [1] doesnot affect the FRE in our experiment, which, if only attenuation of

0 20 40 60 80 100 120 140 160−20

0

20

40

60

80

100

Target Force [N]

Forc

e R

epro

duct

ion

Erro

r [%

]

0 20 40 60 80 100 120 140 160−30

−25

−20

−15

−10

−5

0

5

10

15

20

Target Force [N]

Forc

e R

epro

duct

ion

Erro

r [N

]

* * * *

* * * *

A

B

Figure 3: A, The mean FRE of the group as a function of the targetforce levels. Errorbars: Standard error across subjects. There isan effect of force level on the FRE. The outcomes of the t-tests arepresented in table 2. B, The mean FRE of the group relative to targetforce (FRE/targetforce). Errorbars: Standard error across subjects.The FRE and relative FRE for four force levels differed from zero(stars).

619

self-generated forces causes the FRE, should result in correct forceestimates. However, in this study we found a force level dependentFRE, indicating that attenuation of self-generated forces is not thesole factor in force reproduction tasks.

Walsh et al. [7], demonstrated that the FRE consists of a con-stant component (theoretical FRE at zero Newton) and a force-dependent component (gradient of the slope). He proposed thatthe force-dependent component is due to an enhancement of reaf-ference when receiving the externally applied target force. As-suming that the CNS does not change the sensitivity of the forcesensors between the reference and reproduction trial in this studyand the force-dependency of FRE is caused by an enhancementof reafference in the target (reference) phase, we should find noforce-dependent component. Yet, in our study we did find a force-dependent component, indicating that an enhancement of reaffer-ence is not the cause of the force-dependent component of the FRE.

We suggest that the CNS uses all available information from bothforce sensors and reafference (or efference copy), to estimate theexperienced force using multi-sensory integration. We hypothesizethat the different force sensors are more sensitive, and thus moreaccurate, at different force levels and the accuracy of the sensorschanges over force level. The change in accuracy will bias (under-or overestimating) the force estimate of that sensor. The CNSweights the sensors using their accuracy as proposed in Bayesiandecision theory [4], [8]. If we assume that sensors are unbiased andaccurate at the force levels they are sensitive for and inaccurate andbiased for higher/lower force levels, the total force estimate wouldbe weighted towards the estimate of the accurate sensors, but stillbe biased by the estimate of the inaccurate sensors. The FRE aspreviously found for low force levels [7], [5], can be explained as-suming the sensitivity of the GTO and tactile sensors are for highand low force levels respectively and will bias the force estimates athigher or lower force levels, resulting in too high forces at low forcelevels (GTO) and too low forces at higher force levels (tactile). Atlow force levels, tactile sensors will have an accurate estimate andwill be weighted higher than the inaccurate and biased estimate ofthe GTOs, but the total estimate will be biased towards the GTOestimate. Subjects generated too low forces for higher force levelsand showed an increasing variance with force level when perceiv-ing the externally applied forces using only tactile sensors (Walshet al. 2011, Fig. 5B), which might give an indication of the forcesensitivity of tactile sensors.

Further research is needed to provide a better overview of the be-havior of the force sensors and how the efference copy contributesto the sensory integration process.

Impact for tele-manipulation systems In tele-manipulationsystems, the master and slave transmit forces between the humanoperator and remote environment.This study provides useful infor-mation about the accuracy of the human force reproduction capabil-ities, which can be used to scale forces from and to the human op-erator. The controller might scale the forces from the environmentto force levels, where humans perceive the force more accurately.On the other hand, the forces generated by the human controller canbe corrected to the intended ones. Consequently resulting in moreaccurate and safer tele-manipulation systems.

The relative FRE, Fig. 3B, shows that humans are most sensitiveto forces between 70N and 100N, having correct force estimatesand small variance between subjects. Humans are least sensitive tolow force levels, near 10N, so scaling up low environmental forcesmight improve the human performance. The variance between sub-jects for low forces makes it difficult to accurately correct the forceapplied by the human, but for high force levels this might be a goodoption.

Although we can improve the human performance, further re-search is needed into the effect of force scaling on control effortand mental load.

5 CONCLUSION

In this study we analysed the effect of force level on the force re-production error of humans when generating both the target andreproduction force with the same hand. We found that force levelaffects the force reproduction error and subjects generate higherreproduction forces for force levels up to 40N and lower reproduc-tion forces for force levels of 130N and up. Our results supportforce-dependent sensory integration and show that attenuation ofself-generated forces is not the sole factor in force reproduction er-rors.

ACKNOWLEDGEMENTS

This work is part of the H-Haptic programme from STW. This re-search is supported by the Dutch Technology Foundation STW,which is part of the Netherlands Organization for Scientific Re-search (NWO) and partly funded by the Ministry of Economic Af-fairs, Agriculture and Innovation.

REFERENCES

[1] P. M. Bays and D. M. Wolpert. Computational principles of sensori-motor control that minimize uncertainty and variability. The Journal ofphysiology, 578(Pt 2):387–96, Jan. 2007.

[2] L. A. Jones and I. W. Hunter. Force sensation in isometric contractions:a relative force effect. Brain research, 244(1):186–9, July 1982.

[3] K. Kording. Decision theory: what ”should” the nervous system do?Science, 318(5850):606–10, Oct. 2007.

[4] K. P. Kording and D. M. Wolpert. Bayesian decision theory in sensori-motor control. Trends in cognitive sciences, 10(7):319–26, July 2006.

[5] S. S. Shergill, P. M. Bays, C. D. Frith, and D. M. Wolpert. Two eyes foran eye: the neuroscience of force escalation. Science, 301(5630):187,July 2003.

[6] R. J. van Beers, D. M. Wolpert, and P. Haggard. When feeling is moreimportant than seeing in sensorimotor adaptation. Current biology,12(10):834–7, May 2002.

[7] L. D. Walsh, J. L. Taylor, and S. C. Gandevia. Overestimation of forceduring matching of externally generated forces. Journal of Physiology,3:547–557, 2011.

[8] D. M. Wolpert. Probabilistic models in human sensorimotor control.Human movement science, 26(4):511–24, Aug. 2007.

620